AustingDong
commited on
Commit
·
73c356e
1
Parent(s):
6d117d1
finished
Browse files- app.py +86 -212
- demo/model_utils.py +29 -12
- demo/visualization.py +25 -26
app.py
CHANGED
@@ -22,14 +22,15 @@ def set_seed(model_seed = 42):
|
|
22 |
torch.cuda.manual_seed(model_seed) if torch.cuda.is_available() else None
|
23 |
|
24 |
set_seed()
|
25 |
-
clip_utils = Clip_Utils()
|
26 |
-
clip_utils.init_Clip()
|
27 |
model_utils, vl_gpt, tokenizer = None, None, None
|
28 |
-
|
|
|
|
|
|
|
|
|
29 |
language_model_max_layer = 24
|
30 |
-
language_model_best_layer_min =
|
31 |
-
language_model_best_layer_max =
|
32 |
-
vision_model_best_layer = 24
|
33 |
|
34 |
def clean():
|
35 |
global model_utils, vl_gpt, tokenizer, clip_utils
|
@@ -71,123 +72,83 @@ def multimodal_understanding(model_type,
|
|
71 |
|
72 |
input_text_decoded = ""
|
73 |
answer = ""
|
74 |
-
|
75 |
|
76 |
-
|
|
|
77 |
|
78 |
|
79 |
-
|
80 |
-
# Generate Grad-CAM
|
81 |
-
all_layers = [layer.layer_norm1 for layer in clip_utils.model.vision_model.encoder.layers]
|
82 |
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
cam = cam.to("cpu")
|
90 |
-
cam = [generate_gradcam(cam, image, size=(224, 224))]
|
91 |
-
grad_cam.remove_hooks()
|
92 |
-
target_token_decoded = ""
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
else:
|
97 |
-
|
98 |
-
for param in vl_gpt.parameters():
|
99 |
-
param.requires_grad = True
|
100 |
|
|
|
|
|
|
|
|
|
101 |
|
102 |
-
|
|
|
103 |
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
print("
|
115 |
-
|
116 |
-
|
117 |
-
input_ids_decoded = [tokenizer.decode([input_ids[i]]) for i in range(len(input_ids))]
|
118 |
-
|
119 |
-
if activation_map_method == "GradCAM":
|
120 |
-
# target_layers = vl_gpt.vision_model.vision_tower.blocks
|
121 |
-
if focus == "Visual Encoder":
|
122 |
-
if model_name.split('-')[0] == "Janus":
|
123 |
-
all_layers = [block.norm1 for block in vl_gpt.vision_model.vision_tower.blocks]
|
124 |
-
else:
|
125 |
-
all_layers = [block.layer_norm1 for block in vl_gpt.vision_tower.vision_model.encoder.layers]
|
126 |
-
else:
|
127 |
-
all_layers = [layer.self_attn for layer in vl_gpt.language_model.model.layers]
|
128 |
-
|
129 |
-
print("layer values:", visualization_layer_min, visualization_layer_max)
|
130 |
-
if visualization_layer_min != visualization_layer_max:
|
131 |
-
print("multi layers")
|
132 |
-
target_layers = all_layers[visualization_layer_min-1 : visualization_layer_max]
|
133 |
-
else:
|
134 |
-
print("single layer")
|
135 |
-
target_layers = [all_layers[visualization_layer_min-1]]
|
136 |
-
|
137 |
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
cam = [cam_i]
|
154 |
-
else:
|
155 |
-
i = 0
|
156 |
-
cam = []
|
157 |
-
while start + i < len(input_ids_decoded):
|
158 |
-
if model_name.split('-')[0] == "Janus":
|
159 |
-
gradcam = VisualizationJanus(vl_gpt, target_layers)
|
160 |
-
elif model_name.split('-')[0] == "LLaVA":
|
161 |
-
gradcam = VisualizationLLaVA(vl_gpt, target_layers)
|
162 |
-
elif model_name.split('-')[0] == "ChartGemma":
|
163 |
-
gradcam = VisualizationChartGemma(vl_gpt, target_layers)
|
164 |
-
cam_tensors, grid_size, start = gradcam.generate_cam(prepare_inputs, tokenizer, temperature, top_p, i, visual_method, focus, accumulate_method)
|
165 |
-
cam_grid = cam_tensors.reshape(grid_size, grid_size)
|
166 |
-
cam_i = generate_gradcam(cam_grid, image)
|
167 |
-
cam_i = add_title_to_image(cam_i, input_ids_decoded[start + i])
|
168 |
-
cam.append(cam_i)
|
169 |
-
gradcam.remove_hooks()
|
170 |
-
i += 1
|
171 |
else:
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
|
190 |
-
|
|
|
|
|
|
|
|
|
|
|
191 |
|
192 |
|
193 |
# Collect Results
|
@@ -219,34 +180,7 @@ def model_slider_change(model_type):
|
|
219 |
global model_utils, vl_gpt, tokenizer, clip_utils, model_name, language_model_max_layer, language_model_best_layer_min, language_model_best_layer_max, vision_model_best_layer
|
220 |
model_name = model_type
|
221 |
|
222 |
-
|
223 |
-
encoder_only_res = [
|
224 |
-
gr.Dropdown(choices=["Visualization only"], value="Visualization only", label="response_type"),
|
225 |
-
gr.Dropdown(choices=["Visual Encoder"], value="Visual Encoder", label="focus"),
|
226 |
-
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type"),
|
227 |
-
gr.Dropdown(choices=["CLS", "max", "avg"], value="CLS", label="visual pooling method")
|
228 |
-
]
|
229 |
-
|
230 |
-
language_res = [
|
231 |
-
gr.Dropdown(choices=["Visualization only", "answer + visualization"], value="answer + visualization", label="response_type"),
|
232 |
-
gr.Dropdown(choices=["Language Model"], value="Language Model", label="focus"),
|
233 |
-
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type"),
|
234 |
-
gr.Dropdown(choices=["softmax", "sigmoid"], value="softmax", label="activation function")
|
235 |
-
]
|
236 |
-
|
237 |
-
|
238 |
-
if model_type == "Clip":
|
239 |
-
clean()
|
240 |
-
set_seed()
|
241 |
-
clip_utils = Clip_Utils()
|
242 |
-
clip_utils.init_Clip()
|
243 |
-
sliders = [
|
244 |
-
gr.Slider(minimum=1, maximum=12, value=12, step=1, label="visualization layers min"),
|
245 |
-
gr.Slider(minimum=1, maximum=12, value=12, step=1, label="visualization layers max"),
|
246 |
-
]
|
247 |
-
return tuple(encoder_only_res + sliders)
|
248 |
-
|
249 |
-
elif model_type.split('-')[0] == "Janus":
|
250 |
# best seed: 70
|
251 |
clean()
|
252 |
set_seed()
|
@@ -263,7 +197,7 @@ def model_slider_change(model_type):
|
|
263 |
gr.Slider(minimum=1, maximum=24, value=language_model_best_layer_min, step=1, label="visualization layers min"),
|
264 |
gr.Slider(minimum=1, maximum=24, value=language_model_best_layer_max, step=1, label="visualization layers max"),
|
265 |
]
|
266 |
-
return tuple(
|
267 |
|
268 |
elif model_type.split('-')[0] == "LLaVA":
|
269 |
|
@@ -280,7 +214,7 @@ def model_slider_change(model_type):
|
|
280 |
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer_min, step=1, label="visualization layers min"),
|
281 |
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer_max, step=1, label="visualization layers max"),
|
282 |
]
|
283 |
-
return tuple(
|
284 |
|
285 |
elif model_type.split('-')[0] == "ChartGemma":
|
286 |
clean()
|
@@ -290,62 +224,16 @@ def model_slider_change(model_type):
|
|
290 |
for layer in vl_gpt.language_model.model.layers:
|
291 |
layer.self_attn = ModifiedGemmaAttention(layer.self_attn)
|
292 |
language_model_max_layer = 18
|
293 |
-
|
294 |
-
language_model_best_layer_min = 11
|
295 |
language_model_best_layer_max = 15
|
296 |
|
297 |
sliders = [
|
298 |
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer_min, step=1, label="visualization layers min"),
|
299 |
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer_max, step=1, label="visualization layers max"),
|
300 |
]
|
301 |
-
return tuple(
|
302 |
-
|
303 |
-
|
304 |
|
305 |
|
306 |
-
def focus_change(focus):
|
307 |
-
global model_name, language_model_max_layer
|
308 |
-
if model_name == "Clip":
|
309 |
-
res = (
|
310 |
-
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type"),
|
311 |
-
gr.Slider(minimum=1, maximum=12, value=12, step=1, label="visualization layers min"),
|
312 |
-
gr.Slider(minimum=1, maximum=12, value=12, step=1, label="visualization layers max")
|
313 |
-
)
|
314 |
-
return res
|
315 |
-
|
316 |
-
if focus == "Language Model":
|
317 |
-
if response_type.value == "answer + visualization":
|
318 |
-
res = (
|
319 |
-
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type"),
|
320 |
-
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer_min, step=1, label="visualization layers min"),
|
321 |
-
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer_max, step=1, label="visualization layers max")
|
322 |
-
)
|
323 |
-
return res
|
324 |
-
else:
|
325 |
-
res = (
|
326 |
-
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type"),
|
327 |
-
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer_min, step=1, label="visualization layers min"),
|
328 |
-
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer_max, step=1, label="visualization layers max")
|
329 |
-
)
|
330 |
-
return res
|
331 |
-
|
332 |
-
else:
|
333 |
-
if model_name.split('-')[0] == "ChartGemma":
|
334 |
-
res = (
|
335 |
-
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type"),
|
336 |
-
gr.Slider(minimum=1, maximum=26, value=vision_model_best_layer, step=1, label="visualization layers min"),
|
337 |
-
gr.Slider(minimum=1, maximum=26, value=vision_model_best_layer, step=1, label="visualization layers max")
|
338 |
-
)
|
339 |
-
return res
|
340 |
-
|
341 |
-
else:
|
342 |
-
res = (
|
343 |
-
gr.Dropdown(choices=["GradCAM"], value="GradCAM", label="activation map type"),
|
344 |
-
gr.Slider(minimum=1, maximum=24, value=24, step=1, label="visualization layers min"),
|
345 |
-
gr.Slider(minimum=1, maximum=24, value=24, step=1, label="visualization layers max")
|
346 |
-
)
|
347 |
-
return res
|
348 |
-
|
349 |
|
350 |
def test_change(test_selector):
|
351 |
if test_selector == "mini-VLAT":
|
@@ -376,7 +264,7 @@ with gr.Blocks() as demo:
|
|
376 |
with gr.Row():
|
377 |
|
378 |
with gr.Column():
|
379 |
-
model_selector = gr.Dropdown(choices=["
|
380 |
test_selector = gr.Dropdown(choices=["mini-VLAT", "VLAT", "VLAT-old"], value="mini-VLAT", label="test")
|
381 |
question_input = gr.Textbox(label="Input Prompt")
|
382 |
und_seed_input = gr.Number(label="Seed", precision=0, value=42)
|
@@ -386,15 +274,15 @@ with gr.Blocks() as demo:
|
|
386 |
|
387 |
|
388 |
with gr.Column():
|
389 |
-
response_type = gr.Dropdown(choices=["Visualization only"], value="
|
390 |
-
focus = gr.Dropdown(choices=["
|
391 |
-
activation_map_method = gr.Dropdown(choices=["
|
392 |
accumulate_method = gr.Dropdown(choices=["sum", "mult"], value="sum", label="layers accumulate method")
|
393 |
-
visual_method = gr.Dropdown(choices=["
|
394 |
|
395 |
|
396 |
-
visualization_layers_min = gr.Slider(minimum=1, maximum=
|
397 |
-
visualization_layers_max = gr.Slider(minimum=1, maximum=
|
398 |
|
399 |
|
400 |
|
@@ -404,24 +292,10 @@ with gr.Blocks() as demo:
|
|
404 |
fn=model_slider_change,
|
405 |
inputs=model_selector,
|
406 |
outputs=[
|
407 |
-
response_type,
|
408 |
-
focus,
|
409 |
-
activation_map_method,
|
410 |
-
visual_method,
|
411 |
visualization_layers_min,
|
412 |
visualization_layers_max
|
413 |
]
|
414 |
)
|
415 |
-
|
416 |
-
focus.change(
|
417 |
-
fn = focus_change,
|
418 |
-
inputs = focus,
|
419 |
-
outputs=[
|
420 |
-
activation_map_method,
|
421 |
-
visualization_layers_min,
|
422 |
-
visualization_layers_max,
|
423 |
-
]
|
424 |
-
)
|
425 |
|
426 |
|
427 |
|
|
|
22 |
torch.cuda.manual_seed(model_seed) if torch.cuda.is_available() else None
|
23 |
|
24 |
set_seed()
|
|
|
|
|
25 |
model_utils, vl_gpt, tokenizer = None, None, None
|
26 |
+
model_utils = ChartGemma_Utils()
|
27 |
+
vl_gpt, tokenizer = model_utils.init_ChartGemma()
|
28 |
+
for layer in vl_gpt.language_model.model.layers:
|
29 |
+
layer.self_attn = ModifiedGemmaAttention(layer.self_attn)
|
30 |
+
model_name = "ChartGemma-3B"
|
31 |
language_model_max_layer = 24
|
32 |
+
language_model_best_layer_min = 9
|
33 |
+
language_model_best_layer_max = 15
|
|
|
34 |
|
35 |
def clean():
|
36 |
global model_utils, vl_gpt, tokenizer, clip_utils
|
|
|
72 |
|
73 |
input_text_decoded = ""
|
74 |
answer = ""
|
75 |
+
|
76 |
|
77 |
+
for param in vl_gpt.parameters():
|
78 |
+
param.requires_grad = True
|
79 |
|
80 |
|
81 |
+
prepare_inputs = model_utils.prepare_inputs(question, image)
|
|
|
|
|
82 |
|
83 |
+
if response_type == "answer + visualization":
|
84 |
+
if model_name.split('-')[0] == "Janus":
|
85 |
+
inputs_embeds = model_utils.generate_inputs_embeddings(prepare_inputs)
|
86 |
+
outputs = model_utils.generate_outputs(inputs_embeds, prepare_inputs, temperature, top_p)
|
87 |
+
else:
|
88 |
+
outputs = model_utils.generate_outputs(prepare_inputs, temperature, top_p)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
89 |
|
90 |
+
sequences = outputs.sequences.cpu().tolist()
|
91 |
+
answer = tokenizer.decode(sequences[0], skip_special_tokens=True)
|
92 |
+
attention_raw = outputs.attentions
|
93 |
+
print("answer generated")
|
94 |
|
95 |
+
input_ids = prepare_inputs.input_ids[0].cpu().tolist()
|
96 |
+
input_ids_decoded = [tokenizer.decode([input_ids[i]]) for i in range(len(input_ids))]
|
97 |
|
98 |
+
if activation_map_method == "AG-CAM":
|
99 |
+
# target_layers = vl_gpt.vision_model.vision_tower.blocks
|
100 |
+
|
101 |
+
all_layers = [layer.self_attn for layer in vl_gpt.language_model.model.layers]
|
102 |
+
|
103 |
+
print("layer values:", visualization_layer_min, visualization_layer_max)
|
104 |
+
if visualization_layer_min != visualization_layer_max:
|
105 |
+
print("multi layers")
|
106 |
+
target_layers = all_layers[visualization_layer_min-1 : visualization_layer_max]
|
107 |
+
else:
|
108 |
+
print("single layer")
|
109 |
+
target_layers = [all_layers[visualization_layer_min-1]]
|
110 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
111 |
|
112 |
+
if model_name.split('-')[0] == "Janus":
|
113 |
+
gradcam = VisualizationJanus(vl_gpt, target_layers)
|
114 |
+
elif model_name.split('-')[0] == "LLaVA":
|
115 |
+
gradcam = VisualizationLLaVA(vl_gpt, target_layers)
|
116 |
+
elif model_name.split('-')[0] == "ChartGemma":
|
117 |
+
gradcam = VisualizationChartGemma(vl_gpt, target_layers)
|
118 |
|
119 |
+
start = 0
|
120 |
+
cam = []
|
121 |
+
|
122 |
+
# utilize the entire sequence, including <image>s, question, and answer
|
123 |
+
entire_inputs = prepare_inputs
|
124 |
+
if response_type == "answer + visualization" and focus == "question + answer":
|
125 |
+
if model_name.split('-')[0] == "Janus" or model_name.split('-')[0] == "LLaVA":
|
126 |
+
entire_inputs = model_utils.prepare_inputs(question, image, answer)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
127 |
else:
|
128 |
+
entire_inputs["input_ids"] = outputs.sequences
|
129 |
+
entire_inputs["attention_mask"] = torch.ones_like(outputs.sequences)
|
130 |
+
input_ids = entire_inputs['input_ids'][0].cpu().tolist()
|
131 |
+
input_ids_decoded = [tokenizer.decode([input_ids[i]]) for i in range(len(input_ids))]
|
132 |
+
|
133 |
+
cam_tensors, grid_size, start = gradcam.generate_cam(entire_inputs, tokenizer, temperature, top_p, target_token_idx, visual_method, "Language Model", accumulate_method)
|
134 |
+
if target_token_idx != -1:
|
135 |
+
input_text_decoded = input_ids_decoded[start + target_token_idx]
|
136 |
+
for i, cam_tensor in enumerate(cam_tensors):
|
137 |
+
if i == target_token_idx:
|
138 |
+
cam_grid = cam_tensor.reshape(grid_size, grid_size)
|
139 |
+
cam_i = generate_gradcam(cam_grid, image)
|
140 |
+
cam = [add_title_to_image(cam_i, input_text_decoded)]
|
141 |
+
break
|
142 |
+
else:
|
143 |
+
cam = []
|
144 |
+
for i, cam_tensor in enumerate(cam_tensors):
|
145 |
+
cam_grid = cam_tensor.reshape(grid_size, grid_size)
|
146 |
+
cam_i = generate_gradcam(cam_grid, image)
|
147 |
+
cam_i = add_title_to_image(cam_i, input_ids_decoded[start + i])
|
148 |
+
|
149 |
+
cam.append(cam_i)
|
150 |
+
|
151 |
+
gradcam.remove_hooks()
|
152 |
|
153 |
|
154 |
# Collect Results
|
|
|
180 |
global model_utils, vl_gpt, tokenizer, clip_utils, model_name, language_model_max_layer, language_model_best_layer_min, language_model_best_layer_max, vision_model_best_layer
|
181 |
model_name = model_type
|
182 |
|
183 |
+
if model_type.split('-')[0] == "Janus":
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
184 |
# best seed: 70
|
185 |
clean()
|
186 |
set_seed()
|
|
|
197 |
gr.Slider(minimum=1, maximum=24, value=language_model_best_layer_min, step=1, label="visualization layers min"),
|
198 |
gr.Slider(minimum=1, maximum=24, value=language_model_best_layer_max, step=1, label="visualization layers max"),
|
199 |
]
|
200 |
+
return tuple(sliders)
|
201 |
|
202 |
elif model_type.split('-')[0] == "LLaVA":
|
203 |
|
|
|
214 |
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer_min, step=1, label="visualization layers min"),
|
215 |
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer_max, step=1, label="visualization layers max"),
|
216 |
]
|
217 |
+
return tuple(sliders)
|
218 |
|
219 |
elif model_type.split('-')[0] == "ChartGemma":
|
220 |
clean()
|
|
|
224 |
for layer in vl_gpt.language_model.model.layers:
|
225 |
layer.self_attn = ModifiedGemmaAttention(layer.self_attn)
|
226 |
language_model_max_layer = 18
|
227 |
+
language_model_best_layer_min = 9
|
|
|
228 |
language_model_best_layer_max = 15
|
229 |
|
230 |
sliders = [
|
231 |
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer_min, step=1, label="visualization layers min"),
|
232 |
gr.Slider(minimum=1, maximum=language_model_max_layer, value=language_model_best_layer_max, step=1, label="visualization layers max"),
|
233 |
]
|
234 |
+
return tuple(sliders)
|
|
|
|
|
235 |
|
236 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
237 |
|
238 |
def test_change(test_selector):
|
239 |
if test_selector == "mini-VLAT":
|
|
|
264 |
with gr.Row():
|
265 |
|
266 |
with gr.Column():
|
267 |
+
model_selector = gr.Dropdown(choices=["ChartGemma-3B", "Janus-Pro-1B", "Janus-Pro-7B", "LLaVA-1.5-7B"], value="ChartGemma-3B", label="model")
|
268 |
test_selector = gr.Dropdown(choices=["mini-VLAT", "VLAT", "VLAT-old"], value="mini-VLAT", label="test")
|
269 |
question_input = gr.Textbox(label="Input Prompt")
|
270 |
und_seed_input = gr.Number(label="Seed", precision=0, value=42)
|
|
|
274 |
|
275 |
|
276 |
with gr.Column():
|
277 |
+
response_type = gr.Dropdown(choices=["Visualization only", "answer + visualization"], value="answer + visualization", label="response_type")
|
278 |
+
focus = gr.Dropdown(choices=["question", "question + answer"], value="question + answer", label="focus")
|
279 |
+
activation_map_method = gr.Dropdown(choices=["AG-CAM"], value="AG-CAM", label="visualization type")
|
280 |
accumulate_method = gr.Dropdown(choices=["sum", "mult"], value="sum", label="layers accumulate method")
|
281 |
+
visual_method = gr.Dropdown(choices=["softmax", "sigmoid"], value="softmax", label="activation function")
|
282 |
|
283 |
|
284 |
+
visualization_layers_min = gr.Slider(minimum=1, maximum=18, value=11, step=1, label="visualization layers min")
|
285 |
+
visualization_layers_max = gr.Slider(minimum=1, maximum=18, value=15, step=1, label="visualization layers max")
|
286 |
|
287 |
|
288 |
|
|
|
292 |
fn=model_slider_change,
|
293 |
inputs=model_selector,
|
294 |
outputs=[
|
|
|
|
|
|
|
|
|
295 |
visualization_layers_min,
|
296 |
visualization_layers_max
|
297 |
]
|
298 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
299 |
|
300 |
|
301 |
|
demo/model_utils.py
CHANGED
@@ -74,14 +74,14 @@ class Janus_Utils(Model_Utils):
|
|
74 |
return self.vl_gpt, self.tokenizer
|
75 |
|
76 |
@spaces.GPU(duration=120)
|
77 |
-
def prepare_inputs(self, question, image):
|
78 |
conversation = [
|
79 |
{
|
80 |
"role": "<|User|>",
|
81 |
"content": f"<image_placeholder>\n{question}",
|
82 |
"images": [image],
|
83 |
},
|
84 |
-
{"role": "<|Assistant|>", "content": ""}
|
85 |
]
|
86 |
|
87 |
pil_images = [Image.fromarray(image)]
|
@@ -152,16 +152,33 @@ class LLaVA_Utils(Model_Utils):
|
|
152 |
return self.vl_gpt, self.tokenizer
|
153 |
|
154 |
@spaces.GPU(duration=120)
|
155 |
-
def prepare_inputs(self, question, image):
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
165 |
|
166 |
prompt = self.processor.apply_chat_template(conversation, add_generation_prompt=True)
|
167 |
pil_images = [Image.fromarray(image).resize((384, 384))]
|
|
|
74 |
return self.vl_gpt, self.tokenizer
|
75 |
|
76 |
@spaces.GPU(duration=120)
|
77 |
+
def prepare_inputs(self, question, image, answer=None):
|
78 |
conversation = [
|
79 |
{
|
80 |
"role": "<|User|>",
|
81 |
"content": f"<image_placeholder>\n{question}",
|
82 |
"images": [image],
|
83 |
},
|
84 |
+
{"role": "<|Assistant|>", "content": answer if answer else ""}
|
85 |
]
|
86 |
|
87 |
pil_images = [Image.fromarray(image)]
|
|
|
152 |
return self.vl_gpt, self.tokenizer
|
153 |
|
154 |
@spaces.GPU(duration=120)
|
155 |
+
def prepare_inputs(self, question, image, answer=None):
|
156 |
+
if answer:
|
157 |
+
conversation = [
|
158 |
+
{
|
159 |
+
"role": "user",
|
160 |
+
"content": [
|
161 |
+
{"type": "text", "text": question},
|
162 |
+
{"type": "image"},
|
163 |
+
],
|
164 |
+
},
|
165 |
+
{
|
166 |
+
"role": "assistant",
|
167 |
+
"content": [
|
168 |
+
{"type": "text", "text": answer},
|
169 |
+
],
|
170 |
+
}
|
171 |
+
]
|
172 |
+
else:
|
173 |
+
conversation = [
|
174 |
+
{
|
175 |
+
"role": "user",
|
176 |
+
"content": [
|
177 |
+
{"type": "text", "text": question},
|
178 |
+
{"type": "image"},
|
179 |
+
],
|
180 |
+
},
|
181 |
+
]
|
182 |
|
183 |
prompt = self.processor.apply_chat_template(conversation, add_generation_prompt=True)
|
184 |
pil_images = [Image.fromarray(image).resize((384, 384))]
|
demo/visualization.py
CHANGED
@@ -25,7 +25,7 @@ class Visualization:
|
|
25 |
self.hooks.append(layer.register_backward_hook(self._backward_hook))
|
26 |
|
27 |
def _forward_hook(self, module, input, output):
|
28 |
-
print("forward_hook: self_attn_input: ", input)
|
29 |
self.activations.append(output)
|
30 |
|
31 |
def _backward_hook(self, module, grad_in, grad_out):
|
@@ -42,12 +42,12 @@ class Visualization:
|
|
42 |
layer.get_attn_map = types.MethodType(get_attn_map, layer)
|
43 |
|
44 |
def _forward_activate_hooks(self, module, input, output):
|
45 |
-
print("forward_activate_hool: module: ", module)
|
46 |
-
print("forward_activate_hook: self_attn_input: ", input)
|
47 |
|
48 |
attn_output, attn_weights = output # Unpack outputs
|
49 |
-
print("attn_output shape:", attn_output.shape)
|
50 |
-
print("attn_weights shape:", attn_weights.shape)
|
51 |
module.save_attn_map(attn_weights)
|
52 |
attn_weights.register_hook(module.save_attn_gradients)
|
53 |
|
@@ -137,8 +137,10 @@ class Visualization:
|
|
137 |
|
138 |
grad = F.relu(grad)
|
139 |
|
|
|
140 |
# cam = grad
|
141 |
cam = act * grad # shape: [1, heads, seq_len, seq_len]
|
|
|
142 |
cam = cam.sum(dim=1) # shape: [1, seq_len, seq_len]
|
143 |
cam = cam.to(torch.float32).detach().cpu()
|
144 |
cams.append(cam)
|
@@ -187,7 +189,6 @@ class Visualization:
|
|
187 |
# print("cam_sum shape: ", cam_sum.shape)
|
188 |
num_patches = cam_sum.shape[-1] # Last dimension of CAM output
|
189 |
grid_size = int(num_patches ** 0.5)
|
190 |
-
# print(f"Detected grid size: {grid_size}x{grid_size}")
|
191 |
|
192 |
cam_sum = cam_sum.view(grid_size, grid_size)
|
193 |
if normalize:
|
@@ -207,7 +208,6 @@ class Visualization:
|
|
207 |
|
208 |
num_patches = cam_l_i.shape[-1] # Last dimension of CAM output
|
209 |
grid_size = int(num_patches ** 0.5)
|
210 |
-
# print(f"Detected grid size: {grid_size}x{grid_size}")
|
211 |
|
212 |
# Fix the reshaping step dynamically
|
213 |
cam_reshaped = cam_l_i.view(grid_size, grid_size)
|
@@ -258,7 +258,6 @@ class VisualizationClip(Visualization):
|
|
258 |
|
259 |
@spaces.GPU(duration=120)
|
260 |
def generate_cam(self, input_tensor, target_token_idx=None, visual_method="CLS"):
|
261 |
-
""" Generates Grad-CAM heatmap for ViT. """
|
262 |
self.setup_grads()
|
263 |
# Forward Backward pass
|
264 |
output_full = self.forward_backward(input_tensor, visual_method, target_token_idx)
|
@@ -301,9 +300,13 @@ class VisualizationJanus(Visualization):
|
|
301 |
def forward_backward(self, input_tensor, tokenizer, temperature, top_p, target_token_idx=None, visual_method="softmax", focus="Visual Encoder"):
|
302 |
# Forward
|
303 |
image_embeddings, inputs_embeddings, outputs = self.model(input_tensor, tokenizer, temperature, top_p)
|
304 |
-
|
|
|
305 |
start_idx = 620
|
306 |
self.model.zero_grad()
|
|
|
|
|
|
|
307 |
if focus == "Visual Encoder":
|
308 |
loss = outputs.logits.max(dim=-1).values[0, start_idx + target_token_idx]
|
309 |
loss.backward()
|
@@ -335,11 +338,15 @@ class VisualizationJanus(Visualization):
|
|
335 |
|
336 |
elif focus == "Language Model":
|
337 |
|
338 |
-
cam_sum = self.grad_cam_llm(mean_inside=True)
|
339 |
|
340 |
-
images_seq_mask = input_tensor.images_seq_mask
|
341 |
|
342 |
-
cam_sum_lst, grid_size = self.process_multiple(cam_sum, start_idx, images_seq_mask)
|
|
|
|
|
|
|
|
|
343 |
|
344 |
return cam_sum_lst, grid_size, start_idx
|
345 |
|
@@ -407,15 +414,6 @@ class VisualizationChartGemma(Visualization):
|
|
407 |
self._modify_layers()
|
408 |
self._register_hooks_activations()
|
409 |
|
410 |
-
# def custom_loss(self, start_idx, input_ids, logits):
|
411 |
-
# Q = logits.shape[1]
|
412 |
-
# loss = 0
|
413 |
-
# q = 0
|
414 |
-
# while start_idx + q < Q - 1:
|
415 |
-
# loss += F.cross_entropy(logits[0, start_idx + q], input_ids[0, start_idx + q + 1])
|
416 |
-
# q += 1
|
417 |
-
# return loss
|
418 |
-
|
419 |
|
420 |
def forward_backward(self, inputs, focus, start_idx, target_token_idx, visual_method="softmax"):
|
421 |
outputs_raw = self.model(**inputs, output_hidden_states=True)
|
@@ -429,9 +427,11 @@ class VisualizationChartGemma(Visualization):
|
|
429 |
elif focus == "Language Model":
|
430 |
self.model.zero_grad()
|
431 |
print("logits shape:", outputs_raw.logits.shape)
|
|
|
432 |
if target_token_idx == -1:
|
433 |
-
|
434 |
-
|
|
|
435 |
else:
|
436 |
loss = outputs_raw.logits.max(dim=-1).values[0, start_idx + target_token_idx]
|
437 |
loss.backward()
|
@@ -495,7 +495,7 @@ def generate_gradcam(
|
|
495 |
normalize=False
|
496 |
):
|
497 |
"""
|
498 |
-
Generates a
|
499 |
|
500 |
Parameters:
|
501 |
cam (torch.Tensor): A tensor of shape (C, H, W) representing the
|
@@ -508,9 +508,8 @@ def generate_gradcam(
|
|
508 |
normalize (bool): Whether to normalize the heatmap (default False).
|
509 |
|
510 |
Returns:
|
511 |
-
PIL.Image: The image overlaid with the
|
512 |
"""
|
513 |
-
# print("Generating Grad-CAM with shape:", cam.shape)
|
514 |
|
515 |
if normalize:
|
516 |
cam_min, cam_max = cam.min(), cam.max()
|
|
|
25 |
self.hooks.append(layer.register_backward_hook(self._backward_hook))
|
26 |
|
27 |
def _forward_hook(self, module, input, output):
|
28 |
+
# print("forward_hook: self_attn_input: ", input)
|
29 |
self.activations.append(output)
|
30 |
|
31 |
def _backward_hook(self, module, grad_in, grad_out):
|
|
|
42 |
layer.get_attn_map = types.MethodType(get_attn_map, layer)
|
43 |
|
44 |
def _forward_activate_hooks(self, module, input, output):
|
45 |
+
# print("forward_activate_hool: module: ", module)
|
46 |
+
# print("forward_activate_hook: self_attn_input: ", input)
|
47 |
|
48 |
attn_output, attn_weights = output # Unpack outputs
|
49 |
+
# print("attn_output shape:", attn_output.shape)
|
50 |
+
# print("attn_weights shape:", attn_weights.shape)
|
51 |
module.save_attn_map(attn_weights)
|
52 |
attn_weights.register_hook(module.save_attn_gradients)
|
53 |
|
|
|
137 |
|
138 |
grad = F.relu(grad)
|
139 |
|
140 |
+
# cam = act
|
141 |
# cam = grad
|
142 |
cam = act * grad # shape: [1, heads, seq_len, seq_len]
|
143 |
+
|
144 |
cam = cam.sum(dim=1) # shape: [1, seq_len, seq_len]
|
145 |
cam = cam.to(torch.float32).detach().cpu()
|
146 |
cams.append(cam)
|
|
|
189 |
# print("cam_sum shape: ", cam_sum.shape)
|
190 |
num_patches = cam_sum.shape[-1] # Last dimension of CAM output
|
191 |
grid_size = int(num_patches ** 0.5)
|
|
|
192 |
|
193 |
cam_sum = cam_sum.view(grid_size, grid_size)
|
194 |
if normalize:
|
|
|
208 |
|
209 |
num_patches = cam_l_i.shape[-1] # Last dimension of CAM output
|
210 |
grid_size = int(num_patches ** 0.5)
|
|
|
211 |
|
212 |
# Fix the reshaping step dynamically
|
213 |
cam_reshaped = cam_l_i.view(grid_size, grid_size)
|
|
|
258 |
|
259 |
@spaces.GPU(duration=120)
|
260 |
def generate_cam(self, input_tensor, target_token_idx=None, visual_method="CLS"):
|
|
|
261 |
self.setup_grads()
|
262 |
# Forward Backward pass
|
263 |
output_full = self.forward_backward(input_tensor, visual_method, target_token_idx)
|
|
|
300 |
def forward_backward(self, input_tensor, tokenizer, temperature, top_p, target_token_idx=None, visual_method="softmax", focus="Visual Encoder"):
|
301 |
# Forward
|
302 |
image_embeddings, inputs_embeddings, outputs = self.model(input_tensor, tokenizer, temperature, top_p)
|
303 |
+
print(input_tensor.keys())
|
304 |
+
input_ids = input_tensor["input_ids"]
|
305 |
start_idx = 620
|
306 |
self.model.zero_grad()
|
307 |
+
|
308 |
+
|
309 |
+
|
310 |
if focus == "Visual Encoder":
|
311 |
loss = outputs.logits.max(dim=-1).values[0, start_idx + target_token_idx]
|
312 |
loss.backward()
|
|
|
338 |
|
339 |
elif focus == "Language Model":
|
340 |
|
341 |
+
# cam_sum = self.grad_cam_llm(mean_inside=True)
|
342 |
|
343 |
+
images_seq_mask = input_tensor.images_seq_mask[0].detach().cpu().tolist()
|
344 |
|
345 |
+
# cam_sum_lst, grid_size = self.process_multiple(cam_sum, start_idx, images_seq_mask)
|
346 |
+
|
347 |
+
cams = self.attn_guided_cam()
|
348 |
+
cam_sum_lst, grid_size = self.process_multiple_acc(cams, start_idx, images_seq_mask, accumulate_method=accumulate_method)
|
349 |
+
|
350 |
|
351 |
return cam_sum_lst, grid_size, start_idx
|
352 |
|
|
|
414 |
self._modify_layers()
|
415 |
self._register_hooks_activations()
|
416 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
417 |
|
418 |
def forward_backward(self, inputs, focus, start_idx, target_token_idx, visual_method="softmax"):
|
419 |
outputs_raw = self.model(**inputs, output_hidden_states=True)
|
|
|
427 |
elif focus == "Language Model":
|
428 |
self.model.zero_grad()
|
429 |
print("logits shape:", outputs_raw.logits.shape)
|
430 |
+
print("start_idx:", start_idx)
|
431 |
if target_token_idx == -1:
|
432 |
+
logits_prob = F.softmax(outputs_raw.logits, dim=-1)
|
433 |
+
loss = logits_prob.max(dim=-1).values.sum()
|
434 |
+
|
435 |
else:
|
436 |
loss = outputs_raw.logits.max(dim=-1).values[0, start_idx + target_token_idx]
|
437 |
loss.backward()
|
|
|
495 |
normalize=False
|
496 |
):
|
497 |
"""
|
498 |
+
Generates a heatmap overlay on top of the input image.
|
499 |
|
500 |
Parameters:
|
501 |
cam (torch.Tensor): A tensor of shape (C, H, W) representing the
|
|
|
508 |
normalize (bool): Whether to normalize the heatmap (default False).
|
509 |
|
510 |
Returns:
|
511 |
+
PIL.Image: The image overlaid with the heatmap.
|
512 |
"""
|
|
|
513 |
|
514 |
if normalize:
|
515 |
cam_min, cam_max = cam.min(), cam.max()
|